Memory
The only sense that is never off. Every other sense is intermittent. Memory is continuous.
Every rotation inherits what the last one learned. The difference between a tool and an intelligence.
The retrieval engine underneath is QMD, an open-source hybrid search tool created by Tobi Lutke, CEO of Shopify. Interesting enough to test, and verified to work at the depth this system demands. MIT licensed, commercially clear. That's the pattern. The creative director's job was never to invent every tool. It's to spot the right technology, pressure-test it, and compose it into something that serves the work. This system stands on the work of people who build in the open.
Retain
Analyzes work done and extracts what matters. Decisions made. Patterns discovered. Preferences expressed. Predictions recorded. Confidence scored on extraction.
Recall
Surfaces the right memories for the current context. Ranked by relevance, usefulness history, recency, and project match. Fires automatically when context shifts. The retrieval system trains itself.
Reflect
Consolidates the corpus. Merges duplicates. Detects contradictions. Decays unused memories. Promotes patterns that appear across multiple sessions. Proposes, never acts alone. The editorial gate is not optional.
The machine captures. The human decides what it means.
Memories that prove useful rise. Memories that don't, fade. The ranking is not static. It learns from its own retrieval history.
Surface
Memory recalled at startup, ranked by confidence, recency, semantic match.
↓Use
Referenced during the session. It proves relevant to the work.
↓Validate
Session close marks it as useful. Recall count and usefulness score increment.
↓Rank
Next session, usefulness ratio is higher. Memory ranks above others.
↓Compound
Over time, the most valuable memories float to the top. Noise sinks. The floor rises.
Decision
Choices made, approaches rejected, directions approved. The institutional record of what was decided and why.
Feedback
Corrections and confirmations. Preferences, rules, things to never do again. The calibration of the working relationship.
Insight
Patterns discovered, connections realized, understanding that shifted. The moments where knowledge became structural.
Project
What was built, designed, shipped, deployed. The factual record of implementation across all active work.
Prediction
What the Oracle predicted. Structured findings from each simulation. Confidence levels, segments, objections, convergence patterns.
Outcome
What actually happened. Measured reality against prior predictions. Ground truth is the calibration data that closes the loop.
Reference
Contacts, deadlines, resources. The practical facts that need to be findable.
Audit Resolution
Security fixes, governance gaps resolved. The record of what was caught and addressed.
Predict
Oracle runs a simulation. Memory stores the structured findings: probability, segments, objections, confidence.
↓Recall
Next simulation queries memory: "What have I predicted about this domain? Was I right?" Prior calibration injects as context.
↓Measure
Reality unfolds. Outcomes recorded against predictions. The delta between predicted and actual becomes calibration data.
↓Learn
Confidence updates. Calibration pattern stored. The Oracle at rotation 50 carries the memory of every prediction it ever made.